Method _run

caffe2/python/optimizer.py:1499–1530

(self, net, param_init_net, param_info)

        self.engine = engine

    def _run(self, net, param_init_net, param_info):
        param = param_info.blob
        grad = param_info.grad

        if self.alpha <= 0:
            return

        nz = param_init_net.ConstantFill(
            [param], str(param) + "_ftrl_nz", extra_shape=[2], value=0.0
        )
        self._aux_params.local.append(nz)
        if isinstance(grad, core.GradientSlice):
            grad = self.dedup(net, self.sparse_dedup_aggregator, grad)
            net.SparseFtrl(
                [param, nz, grad.indices, grad.values],
                [param, nz],
                engine=self.engine,
                alpha=self.alpha,
                beta=self.beta,
                lambda1=self.lambda1,
                lambda2=self.lambda2,
            )
        else:
            net.Ftrl(
                [param, nz, grad],
                [param, nz],
                engine=self.engine,
                alpha=self.alpha,
                beta=self.beta,
                lambda1=self.lambda1,
                lambda2=self.lambda2,
            )

    def scale_learning_rate(self, scale):
        self.alpha *= scale